Where ChatGPT Gets Its Information: The 3 Sources That Decide If You're Mentioned
ChatGPT does not look things up the way you might think. It pulls from three different sources, each with its own logic. Which one you live in decides whether the model mentions your name.
5 min read
The wrong mental model
ChatGPT composes answers from three distinct sources, not from a single web search. Each one has different rules for who gets mentioned.
Most people picture ChatGPT as a smart Google. Type a question, it searches the web, it gives back a summary. That picture is wrong, and the wrongness is the reason most experts cannot figure out why they are absent from the answers.
ChatGPT does not search and then summarize. It composes a reply from three distinct sources, often layered together in the same answer. Each source has its own logic, its own update cycle, and its own way of letting your name through. Until you can separate them, the work to get cited stays guesswork.
Here are the three. The rest of this piece is what each one is, what gets you in, and why being present in only one of them is not enough.
Source 1: Training data, the static base
Training data is the model's frozen base. Wikipedia, established media, podcast transcripts, books, and high-trust forums are what survive into it. Get cited there and you get cited everywhere downstream.
The largest source by volume is the one frozen inside the model itself. Before ChatGPT, Claude, or Gemini ever answered a question for a user, they were trained on a massive snapshot of public text. Books, websites, articles, forums, code, scientific papers. The snapshot for the most current models is measured in trillions of tokens.
That snapshot has a knowledge cutoff. The model knows the world up to a date and stops. Cutoffs typically sit months to over a year behind today. After cutoff, the only way new information enters the model is through the next training round.
If your name lives inside the training corpus, the model recognizes you without needing to look anything up. It can describe you, attribute claims to you, and recommend you for prompts that match your topic, all from internal weights. This is the highest-confidence form of citation. It also takes the longest to earn.
What actually ends up in training data is the kind of source that survives crawling and curation: Wikipedia entries, established media articles, podcast transcripts on major platforms, books indexed by Google, GitHub READMEs, Reddit threads above a certain karma threshold, technical documentation. Random LinkedIn posts and ephemeral marketing pages mostly do not.
The leverage in this source is patient and structural. Get into the kinds of places future training rounds will pull from. The work compounds across every model release.
Source 2: Retrieval, live but only when triggered
Retrieval is the engine's live web access. It fires only when the question warrants a real-time fetch. Get fetched and you can be cited within hours of publishing.
Modern AI engines do not stop at training data. Most have web access tools. ChatGPT can browse. Claude can fetch. Perplexity is retrieval-first by design and almost never answers without going to the live web. Gemini blends retrieval into many query types automatically.
When the retrieval tool fires, the engine performs something close to a search-and-read in real time. It fetches a few pages, parses them, and uses the content alongside what the model already knew. The output is a synthesized answer with citations to the pages it just visited.
This is the source where you can earn a mention within hours of publishing. A new article, a fresh page, a podcast that just dropped. The engine sees it on its next retrieval pass and uses it. Speed is real here in a way it is not in training.
But retrieval has a trigger condition. The engine has to decide that the question warrants a live fetch. Some queries always trigger retrieval (current events, prices, schedules, anything time-sensitive). Some never do (general explanations the model already knows). Many sit in between, where retrieval fires only if the user asks for sources or if the engine senses uncertainty.
The leverage is the same SEO foundations you already understand, applied to a slightly different goal. Fast, crawlable, well-structured pages. Schema markup. An llms.txt file that tells AI crawlers what is on your site. Author bylines that are real entities. The shift is from "ranks for keyword" to "gets fetched for the question."
Source 3: User sessions, the invisible layer
User sessions feed reinforcement learning that updates which answers a model trusts. Customers who mention you in their AI conversations train future models to mention you back.
The third source almost never gets discussed because it is the hardest to see. Every conversation a user has with ChatGPT, every Claude session, every Perplexity query becomes a signal. Which answers got positive feedback. Which got challenged. Which got shared. Which the user thanked the model for and which they pushed back on.
These signals do not change the current model in real time. They feed reinforcement learning, fine-tuning, and the next training round. Over time, models update which answers they trust to give and which they hedge.
For an expert who wants to be cited, this source matters more than it looks. If real users keep asking about you and the answers keep getting confirmed, the model becomes more confidently citeable about you in future versions. The reverse holds. If mentions of you provoke pushback, the model gets cautious.
The leverage here is genuinely indirect. You cannot prompt-inject your way into reinforcement signals. What you can do is be the kind of expert that customers naturally bring up in their AI conversations. The same dynamic that grew word of mouth in pre-AI markets now reinforces machine memory. Your customers are your distribution, including in their private ChatGPT sessions.
This is the source most easily neglected and the one with the slowest payoff. It is also the only one no competitor can copy quickly. Real human reinforcement is hard to fake.
Why presence in just one source is not enough
Presence in only one source produces a partial answer. The Entity Gap closes when you show up in at least two of the three: training data, retrieval, and user sessions.
Each of the three sources alone produces a partial answer. Add them together and the model has every reason to mention you. Subtract any one and the gap shows.
Training data without retrieval. The model knows you exist, but cannot talk about your work this quarter. You get cited for last year's positions and ignored for this year's launches.
Retrieval without training data. The engine fetches your page when someone explicitly searches for you. It will not bring you up unprompted, because nothing in its trained weights knows to recommend you.
User sessions without the other two. People have asked about you, but the model has nothing solid to say in return. The reinforcement layer can only confirm what one of the other two sources provided first.
The Entity Gap is the distance between who you actually are as an expert and what the model can verify about you across these three sources together. Closing it is what separates experts who get cited from experts who get overlooked.
This is why the work feels coordinated rather than tactical. SEO alone covers retrieval. PR alone covers training. Customer success alone covers user sessions. None of those silos closes the gap on its own. Rings of Entity is the framework that makes them work as one motion. The next article in this series unpacks the framework. The article after that maps each ring to the technical work on your own domain.
Frequently Asked Questions
Does ChatGPT learn from my conversations with it?
Not in the current session. OpenAI has stated user conversations may be used in future training rounds, subject to user privacy settings. The reinforcement happens at the next training cycle, not instantly. Your conversation today does not change how ChatGPT answers tomorrow, but it can shape how the next model release answers a year from now.
How often is the training data of an AI engine updated?
Major foundation models retrain on a cadence measured in months to over a year. Smaller fine-tunes ship more often. Each model release publishes a "knowledge cutoff" that tells you the freshness window of its training data. After cutoff, only retrieval and user-session signals can introduce new information into answers.
Is Wikipedia really the most weighted source for AI engines?
For the major LLMs, Wikipedia carries disproportionate citation weight relative to its share of public text. The structural reason is that it is curated, cross-referenced, and internally consistent. The model treats consistency between sources as evidence of accuracy, and Wikipedia provides exactly that consistency by design.
What is the difference between training data and retrieval for getting cited?
Training data citations come from sources that existed and were stable when the model was trained. The mention is built into the model itself. Retrieval citations come from sources the engine fetches in real time when answering a specific question. Different work earns each one. Training presence rewards long-term, high-authority placement. Retrieval presence rewards a fast, well-structured site that crawlers can read on the day a user asks.
Can I tell which source ChatGPT used to mention me?
Sometimes. Citations that include a linked URL are almost always retrieval. Mentions without URLs, especially of older or stable facts, are typically training data. User-session influence is invisible by design and cannot be attributed to a specific answer. If you want a clear baseline, run the same prompt with web access on and with web access off. The differences expose which source did the work.